That's a pretty significant boost in hardware specs and performance off the M2070. Nvidia doesn't seem to specify the GPU used, but I wonder if this is the first non-gaming card based on gf110 instead of gf100. The specs seem to lend themselves toward this conclusion, with all 512SP active. I suppose a massive under-clock of gf100 could have allowed them to accomplish this as well, but this seems unlikely based on theoretical FP64/32 performance. There's roughly a 14% increase in shader horsepower going from the M2070 to the M2090, but there's nearly a 30% increase in FP64/32 performance.
It would be absolutely awesome if this is in fact based on gf110, because it also means higher performance Quadro cards could be on the way as well.
[citation][nom]11796pcs[/nom]I know there are different drivers for professional and comsumer based graphics solutions but this M2090 GPU looks like if given the appropriate drivers could crush a 6990. 6GB of DDR5? Holy crap. Is there something special about how professional cards are build vs consumer based that makes them impossible for playing games? Any clarification would be nice.[/citation]
There are both hardware and driver differences between the Tesla and Geforce cards. What defines these differences is the intended purpose for the GPU, which in the case of the Tesla series is almost exclusively GPGPU computing. You'll notice that the Tesla M2000 series doesn't have any monitor output ports because of this, it's meant for installation in rack-mount blade systems and servers.
Even if you could run games off this card, I can guarantee you it wouldn't outperform the HD6990. In fact the GTX580 would probably outperform it without a problem. The hardware is tweaked and geared towards its intended purpose. What makes a great high performance gaming card is very different from what makes a great high performance compute card. Raw clock speeds and mem bandwidth isn't as important, so these are often lower then their Geforce counterparts. Stability and data integrity are paramount. In terms of GPU architecture one of the most significant differences is FP64 performance, which is severely crippled on gf100/gf110 based Geforce cards (1/8 FP32, as opposed to 1/2 on Tesla and Quadro cards).
And as you've already pointed out, mem capacity is much higher. Going from 1.5GB to 6GB wouldn't benefit a Geforce card very much, in fact it would probably result in a performance hit due to the lower mem bandwidth, but it can be extremely important when working with large data sets. There's also ECC support, and a number of other hardware differences, but that basically sums it up.
Software and drivers are completely different. Again, entirely different purpose. No game profiles or any of the things you're used to with gaming cards, so again even if you could run games off a Tesla M2090 it would probably run like crap.
Facebook
Twitter
Instagram